Numerical simulation of the two-locus Wright-Fisher stochastic differential equation with application to approximating transition probability densities

Zhangyi He, Mark Beaumont, Feng YubioRxiv 2020 · DOI

TL;DR

This paper introduces a new numerical method for simulating the complex dynamics of gene frequencies in populations with two linked genes under natural selection. By reformulating the underlying mathematical model and employing an adaptive importance sampling technique, the authors provide a more efficient and accurate way to calculate how gene frequencies change over time, which is crucial for understanding evolution.

What Problem Does This Paper Address?

Understanding how gene frequencies change in a population over time is fundamental to evolutionary biology. The Wright-Fisher diffusion model is a powerful mathematical tool for this, especially when considering natural selection. However, most existing methods for calculating the probability of gene frequencies changing between two time points (the "transition probability density") are limited to scenarios with only one gene (locus). Real-world genetics involves multiple genes that can influence each other through linkage and recombination. Extending these methods to two linked loci has been computationally challenging and prone to inaccuracies, especially when dealing with complex selection pressures. This paper aims to bridge this gap by providing a robust numerical approach for the two-locus Wright-Fisher diffusion.

Methods

The authors tackle the problem by:

  • Reformulating the Stochastic Differential Equation (SDE): They transform the original Wright-Fisher SDE for two linked loci into a "closed form" that is more amenable to direct simulation. This reformulation helps avoid numerical instabilities and boundary issues that plague other methods.
  • Developing a Numerical Scheme: They employ a "Stochastic Taylor scheme," specifically the Euler-Maruyama method, to approximate the solutions to the SDE. This is a standard technique for simulating SDEs, but its application here is tailored to the specific two-locus model.
  • Adaptive Importance Sampling: To efficiently estimate the transition probability density between two observed states, they adapt an importance sampling approach. This technique involves generating simulated "sample paths" of gene frequency changes and weighting them to better estimate the desired probability. Their method is inspired by Fearnhead (2008) but introduces an adaptive mechanism to automatically tune parameters, making it more robust across different scenarios.

Datasets

Not discussed in the paper. The study relies on simulated data generated by their proposed methods.

Key Insights and Ideas

  • The power of reformulation: By cleverly rewriting the SDE, the authors unlock a more direct and stable simulation pathway, bypassing common numerical hurdles.
  • Adaptive sampling is key: The ability to automatically adjust the importance sampling strategy based on the observed states significantly improves the accuracy and efficiency of estimating transition probabilities. This is a crucial step beyond existing methods that require manual parameter tuning.
  • Bridging the gap to multi-locus genetics: This work provides a much-needed computational tool for studying the evolutionary dynamics of linked genes, opening doors for more realistic genetic modeling.

Key Results

The paper demonstrates through extensive simulations that their proposed numerical scheme and adaptive importance sampling approach can:

  • Achieve comparable performance to existing methods (like Fearnhead's) for estimating transition probability densities.
  • Avoid the need for manual tuning of parameters, leading to superior performance across different observed states.
  • Provide a more efficient way to simulate gene frequency trajectories under natural selection at two linked loci.

Limitations and Future Work

  • Computational efficiency in high dimensions: While the method is efficient for two loci, the authors acknowledge that extending it to problems with many more linked loci might suffer from computational efficiency issues.
  • Further improvements in importance sampling: They suggest that constructing even more sophisticated "proposal SDEs" for importance sampling could further enhance performance. This involves better ways to measure the "distance" to the target state and guide the simulations.
  • Exact simulation: The authors note that exploring methods for exact simulation of Wright-Fisher diffusion paths could further reduce bias introduced by numerical schemes.

Summary updated

Abstract

Abstract Over the past decade there has been an increasing focus on the application of the Wright-Fisher diffusion to the inference of natural selection from genetic time series. A key ingredient for modelling the trajectory of gene frequencies through the Wright-Fisher diffusion is its transition probability density function. Recent advances in DNA sequencing techniques have made it possible to monitor genomes in great detail over time, which presents opportunities for investigating natural selection while accounting for genetic recombination and local linkage. However, most existing methods for computing the transition probability density function of the Wright-Fisher diffusion are only applicable to one-locus problems. To address two-locus problems, in this work we propose a novel numerical scheme for the Wright-Fisher stochastic differential equation of population dynamics under natural selection at two linked loci. Our key innovation is that we reformulate the stochastic differential equation in a closed form that is amenable to simulation, which enables us to avoid boundary issues and reduce computational costs. We also propose an adaptive importance sampling approach based on the proposal introduced by Fearnhead (2008) for computing the transition probability density of the Wright-Fisher diffusion between any two observed states. We show through extensive simulation studies that our approach can achieve comparable performance to the method of Fearnhead (2008) but can avoid manually tuning the parameter ρ to deliver superior performance for different observed states.
BibTeX
@misc{he2020numerical,
  author = {Zhangyi He and Mark Beaumont and Feng Yu},
  title = {{Numerical simulation of the two-locus Wright-Fisher stochastic differential equation with application to approximating transition probability densities}},
  year = {2020},
  doi = {10.1101/2020.07.21.213769},
  howpublished = {bioRxiv},
}
Figure